Clinical coding is a task related to clinical billing, aiming at annotating medical reports with codes describing diagnoses and treatments. Recently, automatic coding has become a very active research area for which many models, using neural network-based architectures, have been proposed. Most of these approaches are validated on the MIMIC-III dataset. In this paper, we review the quality of this dataset and propose a new one, and then we experiment with a new classifier based on a budget approach, which aims to facilitate this coding.